NLP Whack-A-Mole: Challenges in Cross-Domain Temporal Expression Extraction

Proceedings of NAACL-HLT 2019, pages 3682–3692 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3682 NLP Whack-A-Mole: Challenges in Cross-Domain Temporal Expression Extraction Amy L. Olex, Luke G. Maffey, and Bridget T. McInnes Department of Computer Science Virginia Commonwealth University 401 East Main Street, Richmond VA 23298 alolex,maffeyl,btmcinnes@vcu.edu Abstract Incorporating domain knowledge is vital in building successful natural language process- ing (NLP) applications. Many times, cross- domain application of a tool results in poor performance as the tool does not account for domain-specific attributes. The clinical do- main is challenging in this aspect due to specialized medical terms and nomenclature, shorthand notation, fragmented text, and a va- riety of writing styles used by different medi- cal units. Temporal resolution is an NLP task that, in general, is domain-agnostic because temporal information is represented using a limited lexicon. However, domain-specific as- pects of temporal resolution are present in clinical texts. Here we explore parsing is- sues that arose when running our system, a tool built on Newswire text, on clinical notes in the THYME corpus. Many parsing issues were straightforward to correct; however, a few code changes resulted in a cascading se- ries of parsing errors that had to be resolved before an improvement in performance was observed, revealing the complexity of tem- poral resolution and rule-based parsing. Our system now outperforms current state-of-the- art systems on the THYME corpus with little change in its performance on Newswire texts. 1 Introduction Temporal resolution is required for comprehend- ing many types of communication, including writ- ten texts. This is especially true in clinical texts as patient narratives revolve around when an event happened, such as when a symptom occurred or the frequency a drug was administered (Lee et al., 2017; Sun et al., 2013b). Understanding the tem- poral component in texts is vital for many NLP systems (Tissot et al., 2015) to accurately interpret a patient narrative (Sun et al., 2013b). Some temporal expressions could be consid- ered domain agnostic as there are limited ways to represent information about time, such as for- matted dates or days of the week. However, there are many lexical variations of these standard to- kens. Additionally, vague temporal expressions, relative times, and event durations require contex- tual or implicit knowledge of the subject area for resolution (Sun et al., 2013b). Clinical texts in- clude all these types of temporal expressions, and also contain domain-specific challenges to tem- poral expression identification and normalization, such as differentiating between dosage and time. Additionally, clinical texts frequently use repeated phrases such as “At this time” that are infrequently used in the general domain. These phrases are vague, relative, and require contextual knowledge of the subject area and the time of events to be re- solved (Sun et al., 2013b). In this work we focus on identification of tem- poral expressions in clinical texts using Chrono– a hybrid system that normalizes temporal ex- pressions into the SCATE Schema (Bethard and Parker, 2016). Originally designed on general …

Content truncated to 500 words